VMware Technical Documentation Embeddings
A specialized sentence-transformers model fine-tuned for semantic search and information retrieval in technical documentation, with a focus on enterprise infrastructure and virtualization technologies.
Model Details
Description
This model extends BAAI/bge-base-en-v1.5 with domain-specific fine-tuning for technical documentation retrieval. It generates 768-dimensional dense embeddings optimized for semantic similarity in enterprise technology contexts.
- Model Type: Sentence Transformer (BERT-based)
- Base Model: BAAI/bge-base-en-v1.5
- Embedding Dimension: 768
- Max Sequence Length: 512 tokens
- Language: English
- License: MIT
Intended Use
Primary Use Cases:
- Semantic search over technical documentation
- Information retrieval for enterprise infrastructure queries
- RAG (Retrieval-Augmented Generation) pipelines
- Technical support knowledge bases
- Enterprise search systems
Optimized For:
- Natural language queries about technical topics
- Documentation retrieval and ranking
- Question answering systems
- Knowledge management platforms
Out-of-Scope
This model is specialized for technical documentation and may not perform optimally for:
- General domain text
- Non-English languages
- Code search or generation
- Creative writing or entertainment content
Quick Start
Installation
pip install sentence-transformers
Basic Usage
from sentence_transformers import SentenceTransformer, util
# Load model
model = SentenceTransformer('BarraHome/vmware-embeddings-large-v1')
# Example queries and documents
queries = [
"How to configure high availability?",
"Steps to install guest tools"
]
documents = [
"High availability can be configured through the management interface...",
"To install guest tools, first mount the ISO image..."
]
# Generate embeddings
query_embeddings = model.encode(queries)
doc_embeddings = model.encode(documents)
# Calculate similarity
similarities = util.cos_sim(query_embeddings, doc_embeddings)
print(similarities)
Semantic Search Example
from sentence_transformers import SentenceTransformer, util
model = SentenceTransformer('BarraHome/vmware-embeddings-large-v1')
# Your document corpus
corpus = [
"Documentation about high availability features...",
"Guide for load balancing configuration...",
"Instructions for live migration procedures..."
]
# Encode corpus
corpus_embeddings = model.encode(corpus, convert_to_tensor=True)
# Query
query = "How to enable high availability?"
query_embedding = model.encode(query, convert_to_tensor=True)
# Search
hits = util.semantic_search(query_embedding, corpus_embeddings, top_k=3)
# Display results
for hit in hits[0]:
print(f"Score: {hit['score']:.4f}")
print(f"Document: {corpus[hit['corpus_id']]}\n")
Performance
Evaluation Metrics
Evaluated on a held-out test set of 2,000 diverse technical queries:
| Metric | Base Model | Fine-tuned | Improvement |
|---|---|---|---|
| Recall@1 | 0.637 | 0.759 | +19.2% |
| Recall@3 | 0.805 | 0.927 | +15.2% |
| Recall@5 | 0.853 | 0.956 | +12.1% |
| Recall@10 | 0.906 | 0.979 | +8.0% |
| NDCG@10 | 0.775 | 0.879 | +13.4% |
Key Performance Indicators
- ✅ 75.9% top-1 accuracy
- ✅ 92.7% top-3 recall
- ✅ 97.9% top-10 recall
- ✅ 0.879 NDCG@10 (excellent ranking quality)
Comparison with Base Model
The fine-tuned model shows consistent improvements across all metrics:
- Higher recall at all k values
- Better ranking quality (NDCG)
- More accurate top-1 predictions
Performance Visualizations
Detailed Metric Comparison:
Percentage Improvements:
Training Details
Training Configuration
- Framework: sentence-transformers
- Loss Function: MultipleNegativesRankingLoss
- Training Strategy: Contrastive learning with hard negative mining
- Epochs: 1
- Batch Size: 64
- Learning Rate: 2e-5 (with 10% warmup)
- Training Samples: 671,972 query-document pairs
- Total Steps: 10,500
- Training Duration: 4 hours 6 minutes
- Throughput: 45.4 samples/second
- Final Loss: 2.245
- Precision: FP16
- Hardware: NVIDIA RTX A6000 (49GB VRAM)
Model Architecture
SentenceTransformer(
(0): Transformer({'max_seq_length': 512, 'do_lower_case': True})
(1): Pooling({'pooling_mode_cls_token': True})
(2): Normalize()
)
Limitations
Known Limitations
- Domain-Specific: Optimized for technical documentation; general domain performance not guaranteed
- English Only: No multi-language support
- Context Length: Limited to 512 tokens
- Recency: Knowledge current as of training date
Recommendations
For optimal results:
-
Query Formulation:
- Use natural language questions
- Include relevant technical terms
- Keep queries under 512 tokens
-
Hybrid Search:
- Combine with keyword search (BM25) for best results
- Use semantic search for understanding, keyword for precision
-
Batch Processing:
- Use
encode(..., batch_size=32)for large collections - Enable
convert_to_tensor=Truefor GPU acceleration
- Use
-
Reranking:
- Consider using a cross-encoder for final reranking
- Retrieve top-100 with this model, rerank to top-10
Technical Specifications
Model Information
- Parameters: ~110M
- Architecture: BERT-base
- Pooling: CLS token
- Normalization: L2
- Similarity Function: Cosine similarity
Performance Benchmarks
| Hardware | Batch Size | Throughput |
|---|---|---|
| RTX 3090 | 32 | ~850 docs/sec |
| A100 | 128 | ~2,100 docs/sec |
| CPU (16 cores) | 8 | ~180 docs/sec |
Resource Requirements
Minimum:
- GPU: 4GB VRAM (batch size 16)
- CPU: 4 cores, 8GB RAM
Recommended:
- GPU: 8GB+ VRAM (batch size 32+)
- CPU: 8+ cores, 16GB+ RAM
Citation
@misc{vmware-embeddings-2024,
author = {Alberto Ferrer},
title = {VMware Technical Documentation Embeddings},
year = {2024},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/BarraHome/vmware-embeddings-large-v1}}
}
Base Model Citation
@misc{bge-base-en-v1.5,
author = {BAAI},
title = {BGE Base English v1.5},
year = {2023},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/BAAI/bge-base-en-v1.5}}
}
Acknowledgments
- Base Model: BAAI/bge-base-en-v1.5 by Beijing Academy of Artificial Intelligence
- Framework: sentence-transformers by UKPLab
License
MIT License
Copyright (c) 2024 [Your Name]
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
Note: This model is intended for research and development. For production use, ensure compliance with your organization's policies and applicable regulations.